Summary
TacCoRL injects tactile feedback into VLA policies without large-scale tactile pretraining or real-world contact exploration. Mixed simulated and real trajectories warm-start tactile-conditioned actions in the pretrained policy, then RL with verifiable task rewards on simulated contact rollouts optimizes near-failure contact behavior — reinforcing tactile-conditioned actions that lead to task completion while a supervised objective on real data keeps the policy anchored to deployment distributions.
Key Contributions
- Scalable two-phase pipeline: warm-start from mixed real+sim tactile trajectories, then RL in a real-aligned simulator targeting near-failure contact states rare in demonstrations
- Avoids both expensive real-world contact exploration (risky and slow) and large-scale tactile-specific pretraining by leveraging simulation for the RL phase exclusively
- Demonstrates that visual VLA priors + targeted tactile RL outperform pure visual or pure tactile approaches on contact-rich manipulation without dedicated tactile datasets
Significance
Shows a practical path to tactile-aware VLAs: simulation handles the contact RL burden while real data anchors the visual-action distribution — enabling contact-rich skill acquisition at the scale and cost of standard VLA fine-tuning.